SENTIMENT ANALYSIS OF IMDB MOVIE REVIEW

DATABASE
A REVIEW PAPER
Avi Ajmera1, Mudit Bhandari2
1, 2
Undergraduate, Department of Computer Engineering
Mukesh Patel School of Technology Management & Engineering
(MPSTME), Shirpur 4 Assistant Professor, Department of Computer
Engineering
Mukesh Patel School of Technology Management & Engineering
(MPSTME), Shirpur

ABSTRACT
Sentiment analysis is an artificial intelligence branch that focuses on
expressing human emotions and opinions in the sense of data. We try to
focus on our sentiment analysis work on the Stanford University-
provided IMDB movie review database in this post. We look at the
expression of emotions to determine whether a movie review is positive
or negative, then standardise and use these features to train multiple
label classifiers to identify movie reviews on the right label. Since there
were no solid framework structures in a study of random jargon movies,
a standardised N-gram approach was used. In addition, a comparison of
the various categories' methods was conducted in order to determine
the best classifier for our domain query. Our process, which employs
separation techniques, has an accuracy of 83 percent.

Keywords: Natural language processing, Sentimental analysis,

lemmatization, Stochastic Gradient Descent (SGD), Support Vector Machine
(SVM), Image Processing

1. INTRODUCTION

““What other people think” has always been the safest way to make well-
informed shopping, service-seeking, and voting decisions. It used to be
confined to friends, relatives, and neighbours; now, thanks to the Internet, it
includes strangers we'd never heard of. On the other hand, an increasing
number of people are using the Internet to share their views with strangers.[1].

Sentiment is a feature that decides whether a text's expressed opinion is

positive or negative, or whether it is about a particular subject. "Logan is a
really nice movie, highly recommended 10/10," for example, sharing positive
feelings about the film Logan, which is also the title of this post. For example,
in the text "I am amazed that so many people put Logan in their favourite
movies, I felt a good watch but not that good," the author's feelings about the
movie might be good, but not the same message. Above everything, [2].
Emotional analysis is a collection of processes, strategies, and tools for
gathering and converting knowledge below the level of language, such as
viewpoint and attitude.

Emotional analysis has traditionally focused on the principle of harmony, or

whether an individual has a positive, neutral, or negative perspective on
something. The topic of emotional analysis was a product or service with
publicly accessible online reviews. This may explain why sentimental analysis
and theoretical analysis are often used interchangeably; however, we believe
that seeing emotions as emotionally charged ideas is more accurate.The aim of
this paper is to categorise movie reviews by examining the magnitude (good or
bad) of each category in the review [3]. We devised an approach to describe the
polarity of the movie using these sentiment words because of the abundant
usage of sentiment words in the study.

To determine which model works best for sentiment classification, we looked

at network configuration and hyper parameters (batch size, number of epochs,
reading rate, batch performance, depth, vocabulary size).

Fig. 1- Discovering People opinions, emotions and feelings about a

product or service

2. METHODOLOGIES
 2.1 Natural language processing (NLP)

Natural language grammar (NLP) is the ability of a computer programme to understand

spoken human language. Artificial intelligence includes NLP. Since computers typically
require people to "speak" to them in the language of an accurate, unambiguous, and well-
structured system, or with a limited number of clearly written voice commands, the creation
of NLP applications is difficult. Human speech, on the other hand, is not always precise; it is
frequently broad, and language systems can be influenced by a variety of factors, including
slang, regional languages, and social context. Natural language processing uses two main
techniques: syntax analysis and semantic naming. Natural language processing uses two main
techniques: syntax analysis and semantic labelling. To understand a method, the syntax word
order in a sentence is essential.

2.2 Sentiment analysis

Sentiment analysis, also known as opinion mining, is a form of natural language processing
(NLP) that decides the emotional tone of a text. This is a common method for businesses to
identify and categorise ideas about a product, service, or concept. It entails mining emotional
text and sensitive knowledge using data analysis, machine learning (ML), and artificial
intelligence (AI)[5]. Emails, blog posts, service tickets, web chats, social
media sites, forums, and tweets are all examples of internet outlets where
technology research tools can help organisations gather insights from
informal and informal messages.
Algorithms may be used to replace manual data analysis with legal,
automatic, or hybrid approaches. We used a few steps in this project to
research and consider the meaning of the term and the context in which it is
used. We began by inserting a summary token and splitting the terms into
features. The random review was stopped in the first phase to ensure that
there was no sexism in the procedure. We kept the movie summary in the
list of positive or bad aspects in front of it until we ended the review. We
then built a new list to store all of the words from the view. The frequency
distribution function was used in the second column to search the most
frequent terms in all updates.

When the most familiar words offer a bad or positive idea, we now
announce the root of these words. We've created a new list of the most
famous terms and their origins, which will be used to train and validate our
method. These terms can now be used as functions of our method to use
various forms of algorithms[6]. We used separate Machine Learning
Algorithms to discern any negative or positive feedback that you have
added to it after separating the training and testing sets.
 2.3 Text Pre-Processing and Normalization
Cleaning, pre-processing, and standardizing the text to bring text artefacts
such as phrases and words to some degree format is one of the most critical
steps before joining the state-of-the-art engineering and modelling process.
This allows for ranking over the entire text, which aids in the creation of
meaningful functionality and reduces the noise that can be imported as a
result of a variety of factors such as inactive characters, special features
characters, XML and HTML tags, and so on.
[9].

Fig. 2 Steps for Text Pre-processing

The following are the key components of our text normalisation pipeline::
Cleaning text: Unnecessary content, such as HTML tags, often appears in
our text and provides little meaning when evaluating emotions. As a
consequence, we must ensure that they are disabled before the functions are
removed. The BeautifulSoup library is doing an excellent job of delivering
the necessary facilities.
Removing accented characters: We work with changes in the English
language in our database, but we need to make sure such characters survive
and that other formats, such as computerised characters, are translated to
ASCII characters..
Expanding contractions: Problems are abbreviated forms of words or
clusters of letters in the English language. By deleting certain letters and
sounds from actual words or phrases, these shortened versions are created.
[10] Typically, vowels are replaced by english.
Removing special characters: Another critical aspect of text processing
and design is the removal of special characters and icons that often add
extra sound to meaningless text.
Stemming and lemmatization: Word titles are always the simplest type of
existing terms and can be created by adding prefixes and suffixes to the
stem to create new words. Inflection is the term for this. A base is the
method of acquiring a simple type of a name on a daily basis.
Lemmatization is similar to stemming in that it eliminates word
conjunctions to expose a word's basic form. The root word, rather than the
root stem, is the basic form in this situation. The distinction is that the root
word is always the proper dictionary word (which is also in the dictionary),
but the root stem may not be.
Removing stopwords: Stopwords or stop words are words that are of little
or no value and are particularly important when creating significant
elements from a text. When you do basic word or word frequency in a copy
text, these are normally the words that end up in the plural.

2.4 Feature weighting methods

 Fig. 3 Feature Weighing Methods

Feature Weighting is Selection of text-related functions depending on the

number of continuous values. We sought to test the efficacy of weight-loss
approaches in the sense of emotional isolation. We also tried to figure out what
impact the split algorithm has. We picked the following three weight-
measurement methods from among the exceptions.

Bag of Words (BoW)

The number of times a word appears in a text is calculated by BoW. The paper
is contrasted with this, and the comparisons are calculated in a variety of
applications. The vector VT Nd is given to it as specified in the text T, with VTi
denoting the number of times the word appears in the text T. D is a lexicon
that includes all of the words in a series of revisions.

TF-IDF
Means a Document Frequency that can be adjusted over time. If it
demonstrates the book's meaning of the expression. This is taken into account
when evaluating diversity mining applications. Increase the number of times a
word appears in the text thus removing the word's occurrence in the corpse.

Word2Vec (W2V)
Term embedding is developed using Word2Vec. This is a model that uses two-
layer neural networks to form word meaning. It accepts a broad corpus of text
as input and generates a space vector of words as output. Each name in the
corpus is assigned a spatial vector [13]. The general meaning in the corpus is
located very similar to the vector space, which is why names are shared. Prior
to running the Word2Vec algorithm, the definitions are separated into
sentences.

2.5 Text classifiers

Text separators are used in sensory research to decide which movie reviews
have the most votes. As a result, we analyse the influence of dividers on
updates using the methods described below.

Random Forest (RF)

It's a way for people to meet. Differentiated integration approaches are used to
solve the approach used to boost performance. A number of vulnerable
students has been grouped together to form a group of good students. The
decision tree is first read by a guarded machine[14].
When we cut under the tree, feedback is given at the apex, and information are
obtained by breaking down the bucket into small sets.

Decision Tree (DT)

A decision tree is a diagram layout that looks like a tree. Each internal hub is
linked to a trait evaluation. The outcome of the test is addressed by each
branch on the hub. The final outcome is discussed by the leaf hub. It is fair in
any new scenario that may occur. It is easy to understand and decode. Other
preference methods can be used to converge a decision tree[15]. The count
becomes perplexing if the attributes are ambiguous or related.

Naïve Bayes (NB)

Based on the Bayes principle. The appearance of one entity in the classroom
has no bearing on the presence of another element. The NB model is
straightforward and effective for massive data sets. The Bayes theorem is used
to measure the back odds.

P(c/x)= (P(c/x)*P(c))/P(x)

P(c/x) is the posterior probability of class given predictor.

P(c) is the prior probability of class.
P(x/c) is the likelihood.
P(x) is the prior probability of predictor
Support Vector Machine (SVM)
It's a machine learning algorithm that's supervised. It is commonly used for
partitioning as well as the withdrawing side. To find a hyper plane, the concept
is to partition the database into two sections. The space between the data point
and the hyper plane is called the boundary. The aim is to assess the absolute
limit of well-distributed new data potential. Smaller collections have higher
accuracy, whereas massive data sharing takes longer.

Neural Network (NN)

They work on a single record at a time. Compares record divisions of
recognised classifications to learn. Errors from the first iteration are fed into
the continuous flow, reducing errors. Back delivery uses forward simulation to
change the weight of the link and is often used to train the network with
proven good performance. It improves the precision of different data. Tuning
and training on a massive database is challenging.

K-Nearest Neighbor (KNN)

It employs all possible segmentation scenarios. The Euclidean parallel scale is
used to establish new classifications. It's used for pattern analysis as well as
statistical equations. The distance between two points in Euclidean space is
known as the square root of the difference between the corresponding point
links[15]. If the complexity of the training database becomes larger, KNN
becomes more complex.

Stochastic Gradient Descent (SGD)

The increasing gradient is another name for it. Upgrade parameter values
based on human training values. This method is useful when an optimization
algorithm is looking for parameters. Instead of renewing coefficients at the
completion of a series of parameters, it is done for each training event[16]. The
gradient drop algorithm makes predictions for each data state, so large data
sets take longer to process. The reduction of stochastic gradient produces a
positive effect of massive data in a short amount of time.
 Fig 4 Classifiers

3 RESULTS AND DISCUSSION

As performance measurements, we used precision, recall, accuracy, and F1

ranking [17].

Fig 5 Performance Metrices

Accuracy
It is a subset of emotions that are distinct from the overall number of
emotions. It's used to work out how effective a machine is.
 Accuracy = (TP+TN)/(TP+TN+FP+FN)

where
TP → number of positive emotions well classified as positive,
FP → positive emotions, but separation does not classify you as good,
TN → number of negative emotions well classified as negative,
FN → negative emotions, but separation does not classify you as bad

Precision
It's a test of how many feelings have been correctly judged as right as opposed
to the overall amount of well-separated emotions (P)

Precision = TP/(TP+FP)

Recall
It is the overall number of sensors in the absolute total that are correctly
predicted (R)

Recall = TP/(TP+FN)
F1 Score
The F measure, which is used to incorporate application accuracy or memory,
is a composite measure that measures a trade's accuracy and memory.

F1=(2*P*R)/(P+R)

o The recall for the SGD classifier was 86.7 percent, while the accuracy,
precision, and F1 ratings for the BoW + SVM combination were 84.4
percent, 83.3 percent, and 85 percent, respectively.

o We obtained high accuracy and precision for SGD classifier by

combining TFIDF with different classifiers (75.1 percent and 78.1
percent, respectively), while TF-IDF + SVM combination provided a
recall of 77.2 percent and F1 score of 75.3 percent.

o We achieved a high precision of 80.4 percent for the RF classifier when

we combined Word2Vec with different classifiers, while Word2Vec +
SGD combination resulted in an accuracy of 82.2 percent, recall of 89.0
 percent, and F1 score of 83.6 percent.

o When we compared precision, we found::

● In terms of BoW, SVM is very strong (83.3 percent )

● In TF-IDF, SGD has a high percentage (80.9%).
● In Word2Vec, Random Forest has a high score (76.6%).

Table 1:Different classifiers with bag of words method

Table 2:Performance of TF-IDF feature weighting method with

different classifiers
 Table 3:On the basis of accuracy, a text classifier is compared to
attribute weighting methods.

Fig 6:Comparison of text classifier with feature weighting methods

based on precision

Table 4:Precision metrics for different machine learning

algorithms are compared.

Fig 7:Focused on the F1 score, a text classifier is compared to

feature weighting methods.
Table 5:F1 score metrics on different machine learning algorithms
compared

4 CONSTRAINTS

The below are the key obstacles for sentiment analysis: -

● Entity With a Name Recognize - What exactly is the person talking
about, for example, is 300 Spartans a group of Greeks or a film?[15]
● Anaphora Resolution -The difficulty of determining what a pronoun
or a noun phrase corresponds to is known as anaphora resolution. "We
went to dinner and watched the movie; it was terrible." What exactly
does "It" mean? [18]
● Parsing - What is the sentence's subject and object, and to whom does
the noun and/or adjective refer?
● Sarcasm -If you don't know the author, you won't know whether "bad"
means "bad" or "good."
● Twitter - Abbreviations, capitalization, bad pronunciation, punctuation,
and syntax are all problems on Twitter.

5 CONCLUSION
This paper examines the issue of sentiment analysis.

There are three influencing factors. The BoW, TF-IDF, and Word2Vec
calculation methods were combined with a variety of classifier, including the
Naive Bayes classifier, Support vector machine, decision tree, Random Forest,
Decreased stochastic gradient, and checks, to determine emotions in the
IMDB's database of movie reviews.
We believe that Word2Vec with SGD is the right mixture of emotions for the
IMDB database segmentation issue based on test results.
The accuracy score was 89 percent, and the F1 score was 83.6 percent, making
this form the best of all combinations. This work can be extended to include
the study of online feedback from social media platforms as well as other
sophisticated algorithms.

REFRENCES
1. A. L. Maas, R.E. Daly, D. Huang, and Ch. Potts, “Learning Word Vectors
for Sentiment Analysis”, Proceedings of 49th
Annual Meeting of the Association for Computational Linguistics:
Human Language Technologies - Volume 1, pp. 142–150, June, 2011

2. Study Of Feature Weighting Techniques For URL Based Webpage

Classification”, S. Wang and C. D. Manning, “Baselines and Bigrams:
Simple, Good Sentiment and Topic Classification”, Proceedings of 50th
Annual Meeting of the Association for Computational Linguistics: Short
Papers - Volume 2, pp. 90–94, July 08–14, 2012.

3. V.K. Singh, P. Waila and R. Piryani “Sentiment Analysis of Movie

Reviews: A new feature-based heuristic for aspect-level sentiment
classification”, International Mutli-Conference on Automation,
Computing, Communication, Control and Compressed Sensing pp. 712–
717, 22–23 March 2013

4. Jiwei Li, “Feature Weight Tuning for Recursive Neural Networks”,

Neural and Evolutionary Computing, pp. 1–11, 13 Dec 2014.

5. Raj K. Palkar, K. D. Gala and Jay N. Shah, “Comparative Evaluation of

Supervised Learning Algorithms for Sentiment Analysis of Movie
Reviews”, International Journal of Computer Applications, Volume 142
– No.1, pp. 20–26, May 2016

6. K. Topal and G. Ozsoyoglu, “Movie Review Analysis: Emotion Analysis of

IMDB Movie Reviews”, IEEE/ACM ASONAM, pp. 1170–1176, 8–21 Aug.
2016
7. R. Rajalakshmi, 2015, “Identifying Health domain URLs using SVM”, In:
Proceedings of the Third International Symposium for Women in
Computing and Informatics, WCI 15, pp. 203– 208, ACM.

8. R. Rajalakshmi and Sanju Xavier, 2017, “Experimental Procedia

Computer Science, Volume 115, pp. 218–225, Elsevier.

9. A. Sajeevan and L. K. S, "An enhanced approach for movie review

analysis using deep learning techniques," 2019 International Conference
on Communication and Electronics Systems (ICCES), Coimbatore, India,
2019, pp. 1788-1794.

10. T. M. Untawale and G. Choudhari, "Implementation of Sentiment

Classification of Movie Reviews by Supervised Machine Learning
Approaches," 2019 3rd International Conference on Computing
Methodologies and Communication (ICCMC), Erode, India, 2019, pp.
1197-1200.

11. M. Yasen and S. Tedmori, "Movies Reviews Sentiment Analysis and

Classification," 2019 IEEE Jordan International Joint Conference on
Electrical Engineering and Information Technology (JEEIT), Amman,
Jordan, 2019, pp. 860-865.

12.S. C. Rachiraju and M. Revanth, "Feature Extraction and Classification of

Movie Reviews using Advanced Machine Learning Models," 2020 4th
International Conference on Intelligent Computing and Control Systems
(ICICCS), Madurai, India, 2020, pp. 814-817.

13.S. Tripathi, R. Mehrotra, V. Bansal and S. Upadhyay, "Analyzing

Sentiment using IMDb Dataset," 2020 12th International Conference on
Computational Intelligence and Communication Networks (CICN),
Bhimtal, India, 2020, pp. 30-33.
14.S. M. Qaisar, "Sentiment Analysis of IMDb Movie Reviews Using Long
Short-Term Memory," 2020 2nd International Conference on Computer
and Information Sciences (ICCIS), Sakaka, Saudi Arabia, 2020, pp. 1-4.

15.M. R. Haque, S. Akter Lima and S. Z. Mishu, "Performance Analysis of

Different Neural Networks for Sentiment Analysis on IMDb Movie
Reviews," 2019 3rd International Conference on Electrical, Computer &
Telecommunication Engineering (ICECTE), Rajshahi, Bangladesh, 2019,
pp. 161-164.

16.K. Sailunaz, M. Dhaliwal, J. Rokne and R. Alhajj, "Emotion detection

from text and speech: a survey", Social Network Analysis and Mining,
vol. 8, no. 1, pp. 28, 2018.

17.B. M. and V. B., "Sentiment Analysis using Support Vector Machine

based on Feature Selection and Semantic Analysis", International
Journal of Computer Applications, vol. 146, no. 13, pp. 26-30, 2016.

18. B. Srinivasa-Desikan, Natural Language Processing and

Computational Linguistics: A practical guide to text analysis with Python
Gensim spaCy and Keras, Packt Publishing Ltd., 2018.